"Combinations of the training data" is also hiding a pretty common and serious misconception about modern generative AI: this seems to come from a mental model

"Combinations of the training data" is also hiding a pretty common and serious misconception about modern generative AI: this seems to come from a mental model that LLMs and image generators are merely producing a collage of sorts, especially that this is some way linked to specific parts of the training data as opposed to a generalized result of building up an understanding of how to produce a sentence, piece of code, or plausible image.